Q-ReAlign scores the perceptual quality / aesthetic appeal of an image or video the way Q-Align does: the model is asked to rate quality, and the probability mass it places on the discrete words excellent / good / fair / poor / bad is collapsed — via a fixed weighting [1.0, 0.75, 0.5, 0.25, 0.0] — into a single scalar in [0, 1].

Lite (4B) is the middle of three sizes (Mini 0.8B · Lite 4B · Pro 9B) and the recommended default: it matches or beats the original Q-Align across the seven QA benchmarks while staying easy to serve.

Backbone: Qwen3.5-VL (model_type: qwen3_5), hybrid linear/full attention text tower + SigLIP-style vision encoder
Tasks: IQA (image quality) · IAA (image aesthetics) · VQA (video quality) — the unified ONE-ALIGN setting
Training: full-parameter SFT in bf16 via ms-swift, vision tower + projector trainable
Precision: bfloat16 · dtype auto

Results

Per-dataset SRCC / PLCC on seven QA benchmarks. Lite (4B) reaches avg SRCC 0.889 vs. Q-Align's 0.869.

Model	KonIQ	SPAQ	KADID	AGI	LIVE	AVA	LSVQ	Avg.
Q-Align	0.942 / 0.944	0.932 / 0.933	0.912 / 0.920	0.738 / 0.781	0.897 / 0.870	0.798 / 0.796	0.867 / 0.866	0.869 / 0.873
Lite (4B)	0.943 / 0.941	0.932 / 0.934	0.928 / 0.931	0.829 / 0.871	0.899 / 0.862	0.814 / 0.804	0.880 / 0.879	0.889 / 0.889

_{Each cell is SRCC / PLCC, on the full evaluation sets (KonIQ, SPAQ, KADID, AGI, LIVE, AVA, LSVQ).}

Quick start

python

import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor
# transformers >= 5.2.0 for Qwen3.5 support

CKPT, IMAGE = "q-future/Q-ReAlign-Lite-4B", "photo.jpg"
LEVELS  = ["excellent", "good", "fair", "poor", "bad"]
WEIGHTS = [1.0, 0.75, 0.5, 0.25, 0.0]

device = "cuda" if torch.cuda.is_available() else "cpu"
processor = AutoProcessor.from_pretrained(CKPT)
model = AutoModelForImageTextToText.from_pretrained(CKPT, dtype="auto").to(device).eval()

messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "How would you rate the quality of this image?"},
]}]
text = processor.apply_chat_template(messages, add_generation_prompt=True) + "The quality of the image is"
inputs = processor(text=[text], images=[Image.open(IMAGE).convert("RGB")], return_tensors="pt").to(device)

ids = [processor.tokenizer(" " + w, add_special_tokens=False).input_ids[0] for w in LEVELS]
probs = model(**inputs).logits[0, -1, ids].softmax(-1)
score = (probs * torch.tensor(WEIGHTS, device=device)).sum().item()
print(f"quality score: {score:.4f}")   # 0 (worst) .. 1 (best)

The score is the expected value of the level weights under the model's next-token distribution over the five level words — no sampling, one forward pass.

Aesthetics or video

Swap the prompt for the task:

Aesthetics (IAA): "How would you rate the aesthetics of this image?" → stem "The aesthetics of the image is"
Video (VQA): sample N frames (default 8) and pass them as the image sequence; prompt "How would you rate the quality of this video?" → stem "The quality of the video is"

Model details

	Lite (4B)
Architecture	`Qwen3_5ForConditionalGeneration`
Text hidden size	2560
Text layers	32 (linear attention with full-attention every 4th layer)
Vision encoder depth	24, hidden 1024, patch 16, spatial merge 2
Vocab	248320
Context length	up to 262144
Tied embeddings	yes
Tensor dtype	bfloat16
Shards	3 × safetensors (~10.3 GB total)

Scoring contract

Level vocabulary: excellent, good, fair, poor, bad
Weights: [1.0, 0.75, 0.5, 0.25, 0.0]
Output: scalar in [0, 1], higher = better
The five level tokens are matched with a leading space (" excellent", …); keep that when porting to other tokenizers.

Intended use & limitations

Use: no-reference image/video quality assessment, aesthetic scoring, dataset curation, ranking and filtering generated media, reward signals for generative pipelines.
Out of scope: safety/content moderation, factual or identity judgments, medical/forensic grading. Quality is perceptual and dataset-conditioned.
Scores are calibrated to the training MOS distribution; absolute values are most meaningful relative to one another. Re-calibrate before mixing with other scales.

Acknowledgements & citation

Built on the shoulders of Q-Align (the discrete text-defined-levels method and ONE-ALIGN), ms-swift (training/inference backbone), and Qwen3.5-VL (the vision-language backbone). If you use this model, please also cite the originals:

bibtex

@inproceedings{wu2024qalign,
  title     = {Q-Align: Teaching {LMM}s for Visual Scoring via Discrete Text-Defined Levels},
  author    = {Wu, Haoning and Zhang, Zicheng and Zhang, Weixia and Chen, Chaofeng and
               Liao, Liang and Li, Chunyi and Gao, Yixuan and Wang, Annan and Zhang, Erli and
               Sun, Wenxiu and Yan, Qiong and Min, Xiongkuo and Zhai, Guangtao and Lin, Weisi},
  booktitle = {Proceedings of the 41st International Conference on Machine Learning (ICML)},
  year      = {2024}
}

@inproceedings{swift2025,
  title     = {{SWIFT}: A Scalable Lightweight Infrastructure for Fine-Tuning},
  author    = {ModelScope Team},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
  year      = {2025},
  note      = {\url{https://github.com/modelscope/ms-swift}}
}

@misc{qwen3_5,
  title        = {Qwen3.5: Towards Native Multimodal Agents},
  author       = {Qwen Team},
  year         = {2025},
  howpublished = {\url{https://github.com/QwenLM/Qwen3-VL}}
}